Self Reported Disability
and Reference Groups
ARTHUR VAN SOESTARIE KAPTEYN TATIANA ANDREYEVA JAMES P. SMITH WR-409-1 November 2006
W O R K I N G
P A P E R
This product is part of the RAND Labor and Population working paper series. RAND working papers are intended to share researchers’ latest findings and to solicit informal peer review. They have been approved for circulation by RAND Labor and Population but have not been formally edited or peer reviewed. Unless otherwise indicated, working papers can be quoted and cited without permission of the author, provided the source is clearly referred to as a working paper. RAND’s publications do not necessarily reflect the opinions of its research clients and sponsors.
Self Reported Disability and Reference Groups
Arthur van Soest, RAND and Tilburg University
yTatiana Andreyeva, Yale University and RAND
Arie Kapteyn, RAND
James P. Smith, RAND
November 6, 2006
Abstract
Individuals are in‡uenced by the types of people with whom they asso-ciate and who form their social networks. These social interactions may a¤ect individual and social norms. We develop a direct test of this using Dutch survey data on how respondents evaluate work disability of hypothetical peo-ple with some work related health problem (vignettes). We analyze how the thresholds respondents use to decide what constitutes a (mild or more serious) work disability depends on the number of people receiving disability insurance bene…ts (DI) in their reference group. To account for endogeneity of DI re-ceipt in a respondent’s reference group, we jointly model the respondent’s own self-reported work disability, the evaluation thresholds, and DI receipt in the reference group. We …nd that reference group e¤ects are signi…cant, and con-tribute substantially to an explanation of why self-reported work disability in the Netherlands is much higher than in, e.g., the US. This implies an important role for social interactions and norms on the perception of work limitations.
This research was funded by the National Institute of Aging.
yRAND, Main Street 1776, Santa Monica CA 90401-3208; email:vansoest@rand.org
1
Introduction
In contrast to other social scientists economists have long adhered to an individu-alistic notion of behavior, despite early contributions by, for example, Duesenberry (1949) and Veblen (1899). An important modern contribution to the modeling of social interactions is the seminal work of Becker (1974). Although of wider relevance, Becker’s work emphasized the interactions among family members, caused by inter-dependent utilities as well as a common budget constraint. In more recent years, economists have increasingly recognized that individual actions are fundamentally in‡uenced by the attributes and behaviors of those other individuals who form their social networks; see Topa (2001).
The span of behaviors that have been examined in this new research on social interactions has been expanding rapidly and even a very partial list now includes criminal activity (Glaeser, Sacerdote & Scheinkman 1996), (Glaeser, Sacerdote & Scheinkman 2000) neighborhood e¤ects on youth behavior (Case & Katz 1991), models of herd or copycat like behaviors (Banerjee 1992), ’peer e¤ects’in education (Hanuschek, Kian, Markman & Rifkin 2000), (Ginther, Haveman & Wolfe 2000), ag-glomeration economies (Audretsch & Feldman 1996), information exchanges in local labor markets (Topa 2001), labor supply (Woittiez & Kapteyn 1998), consumption (Kapteyn, van de Geer, van de Stadt & Wansbeek 1997) (Alessie & Kapteyn 1991), retirement plan choices (Du‡o & Saez 2003) and social learning through neighbors (Bala & Goyal 1998). As these examples illustrate, the type of social interactions studied has moved well beyond the immediate family to much larger circles of friends, neighbors, and like minded consumers and workers. Various reasons are given for why these types of social interactions matter, including information sharing, demon-stration e¤ects, and the formation of tastes and preferences.
Social interactions may also a¤ect what individuals believe to constitute accept-able or normal behavior based on the standards of the sub-communities in which they live and work. In this paper, we develop a direct test of this using data from a household survey representative of the Dutch population on how respondents eval-uate work disability of hypothetical people with some work related health problem (vignettes). Combining this with self-reports on the number of people receiving dis-ability insurance bene…ts (DI) among one’s friends and acquaintances, we estimate a model describing the in‡uence of DI prevalence in one’s reference group on the subjective scale used to report own and others’work disability.
States. Bound & Burkhauser (1999) report that in 1995, the number of DI recipients per 1000 workers in the age group 45-59 was 103 in the U.S., compared to 271 in The Netherlands. Kapteyn, Smith & van Soest (2006) report that in the age bracket 51-64 self reported work disability in The Netherlands is about 58% higher than in the United States (35.8% in The Netherlands against 22.7% in the U.S.). While higher levels of Dutch participation in DI programs is not surprising given higher DI bene…ts and easier eligibility compared to the US1, greater Dutch prevalence of
self-reported work disability is puzzling as the Dutch population appears to be healthier than the Americans. 2
Kapteyn et al. (2006) investigated to what extent di¤erences in self reported work disability can be ascribed to di¤erences in reporting styles and thresholds across coun-tries. Exploiting the vignette methodology originally developed by King, Murray, Sa-lomon & Tandon (2004), Dutch and US respondents were given the same descriptions of work disability problems for hypothetical persons"("vignettes"). Dutch respon-dents appeared to be much more likely to describe the same work disability problem as constituting a work disability than American respondents. Kapteyn et al. (2006) found that more than half of the observed di¤erence in self-reported work disability between the two countries can be explained by di¤erences in response scales.
This result implies that US and Dutch respondents have di¤erent norms for eval-uating work disability. Our paper analyzes to what extent this is due to peer group e¤ects: respondents with many DI recipients in their peer group have social norms making them more likely to evaluate given health problems as constituting a work disability.
We formalize this notion by introducing the concept of prevalence of DI bene…t receipt in one’s reference group, de…ned as one’s circle of friends and acquaintances. In a Dutch survey that we designed and implemented, we asked respondents directly how many people among their friends and acquaintances receive DI bene…ts. In this paper, we develop a model that jointly explains the categorical answer to this question and self-reported work disability. The main feature of the model is the notion that response scales for reporting no, mild, or severe work disability, can be a¤ected by a "peer group e¤ect," i.e., by the number of people in the reference group receiving disability bene…ts. To identify the determinants of response scales,
1See for instance Aarts, Burkhauser & de Jong (1996). In 2004, DI recipients in The Netherlands
made up 13% of the labor force (Source: Statistics Netherlands http://statline.cbs.nl/StatWeb.), while in the US DI-recipients constituted 4.8% of the civilian labor force (Source: US Bureau of Labor Statistics ftp://ftp.bls.gov/pub/news.release/History/empsit.01072005.news)
2This is suggested by the analysis of a broad set of health conditions by Banks, Kapteyn, Smith
& van Soest (2005).
we exploit anchoring vignettes as in Kapteyn et al. (2006). Using this additional information helps to solve the identi…cation problem that is present in many models with peer group e¤ects, known as the re‡ection problem (Manski (1993)).
The remainder of the paper is organized as follows. In the next section we brie‡y describe the micro-data used in our analysis. Section 3 presents the model, which essentially consists of three equations. One equation explains the answers to the question about DI bene…t receipt in the respondents’ reference group. A second equation models self-reported work disability. The third equation (or rather set of equations) explains how individual response scales to questions on work disability (or anchoring vignettes) are a¤ected by the prevalence of DI bene…t receipt in the reference group. Throughout we control for a large number of other variables, such as socio-demographic characteristics and health conditions.
Section 4 summarizes our main results. We …nd that DI bene…t receipt in one’s reference group has a signi…cant e¤ect on response scales in the expected direction. To gauge the size of this e¤ect, we graph the relation between DI bene…t receipt in the reference group against self-reported work disability. It turns out that to explain the complete di¤erence in response scales between the U.S. and The Netherlands, the percentage of respondents in The Netherlands reporting to know at least some DI bene…t recipients has to fall by about half. This is an order of magnitude that seems reasonable given the substantial di¤erence in the number of Dutch and U.S. people on DI bene…ts. The …nal section presents our conclusions.
2
The Data
In this research, we use information obtained from the Dutch CentERpanel. This is an Internet panel of about 2,250 households who have agreed to respond to a survey every weekend. Respondents are recruited by telephone. If they agree to participate and do not already have Internet access, they are provided with Internet access (and if necessary, a set-top box). Thus, the CentERpanel is not restricted to households with Internet access, but representative of the Dutch adult popula-tion except the institupopula-tionalized. Sample weights based upon data from Statistics Netherlands are used to correct for unit nonresponse. The sample that we use to estimate our model consists of about 2,000 respondents who participated in several interviews with questions on work disability in 2003.
vignette evaluations (described below). In October 2003, we …elded a second wave of vignettes with slightly di¤erent wording of the questions, and also included questions about reference groups. For our analysis we will use the vignette and reference group data from this October wave. Appendix A lists the vignette questions. All vignettes are presented with either a female or a male name.3
For each of the vignettes the respondent is asked the following question:
“Does . . . have a health problem that limits the amount or type of work he/she can do?”
with a …ve point response scale:
not at all; yes, mildly limited; yes moderately limited; yes, severely limited; yes, extremely limited/cannot work.
Table 1 presents the response frequencies for each of the 15 vignette questions. The di¤erences in distributions of answers correspond quite well with the variation in severity of the conditions described in the vignettes. For example, in all three domains of a¤ect, pain, and CVD, the condition described in the third vignette seems much more severe than that described in the …rst, and respondents ranked them accordingly. Moreover, there was also a great deal of consistency among respondents in how they ordered vignettes in terms of their severity, showing that respondents understood these experiments and took their responses seriously. (See Banks et al. (2005).
Table 2 presents the distribution of the answers to the question on own work limitations by age group. These represent answers to the question:
"Do you have an impairment or health problem that limits you in the amount or kind of work you can do?".
The question allows respondents to reply on the …ve-point scale (1) No, not at all, (2) Yes, I am somewhat limited, (3) Yes, I am rather limited, (4) Yes, I am severely limited, (5) Yes, I am very severely limited-I am unable to work. These response categories are identical to the ones used to gauge the severity of the vignette work limitations.
Table 2 implies that about 37% of the Dutch population reports to have at least a mild work limitation and about 14% have a work limiting health problem or im-pairment that they gauge as moderately limiting or worse. Not surprisingly, work related health deteriorates with age (although cohort e¤ects may also play some role in this pattern).
3Female or male names are assigned randomly. In Appendix A we only show one of the two
names per vignette.
Table 1. Frequencies for Vignette Answers (CentERpanel, October 2003)
Affect vignettes Affect 1 Affect 2 Affect 3 Affect 4 Affect 5
Not at all limited 41.2 96.2 11.1 18.7 2.2
Somewhat limited 49.7 2.8 44.3 44.8 8.4
Moderately limited 7.2 0.6 31.2 26.0 18.6
Severely limited 1.4 0.5 12.2 8.9 40.4
Extremely limited/cannot work 0.5 0.0 1.3 1.6 30.4
Pain vignettes Pain 1 Pain 2 Pain 3 Pain 4 Pain 5
Not at all limited 22.5 8.2 0.6 0.3 0.8
Somewhat limited 61.8 47.1 6.6 6.2 12.9
Moderately limited 13.4 34.1 25.7 29.4 31.3
Severely limited 1.9 9.2 49.5 43.2 39.2
Extremely limited/cannot work 0.4 1.4 17.6 20.9 15.9
CVD vignettes CVD 1 CVD 2 CVD 3 CVD 4 CVD 5
Not at all limited 91.2 10.6 1.8 20.7 6.7
Somewhat limited 7.8 46.2 18.2 44.9 34.1
Moderately limited 0.9 29.2 32.6 25.0 30.3
Severely limited 0.1 11.8 33.6 8.8 20.7
Extremely limited/cannot work 0.0 2.3 13.9 0.6 8.3
See Appendix 1 for the wordings of the vignette questions. Notes: Data are weighted. Complete sample N=1980.
Table 2. Distribution of Self-Reported Work Disability by Age, %
15-24 25-34 35-44 45-54 55-64 65+ Total
Not at all limited 86.8 74.1 69.2 55.9 52.8 48.4 63.1
Somewhat limited 5.4 20.7 17.5 24.2 28.5 34.3 22.8
Moderately limited 5.8 3.2 5.8 7.0 10.5 10.9 7.1
Severely limited 2.0 0 2.1 2.9 1.8 3.7 2.2
Extremely limited/cannot work 0 1.8 5.4 9.9 6.3 2.8 4.8
Number of observations 68 362 438 460 336 316 1980
Notes: Data are weighted. Complete sample N=1980.
Age Group
Appendix B presents some of the questions about reference groups asked in the October wave and used in the empirical analysis. Our operationalization of a ref-erence group is the circle of acquaintances mentioned in these questions. The …rst two reference group questions provide information on the modal age and modal ed-ucation level in the respondent’s reference group. In the analysis we will combine the age and education categories into a smaller number of broader brackets. Table 3 presents descriptive statistics for our independent variables, including the responses to the …rst two reference group questions listed in Appendix B. For example, 27 percent of all respondents report that most of the people in their reference group are in the age group 36-45. About 48 percent say that most of their acquaintances have medium education level (while 39 percent of the respondents has that level).
The other reference group questions refer to the number of acquaintances receiv-ing disability bene…ts, separately for men and women. These are the crucial variables for our analysis as they measure DI bene…t receipt in the reference group. For men, we will use the number of male acquaintances on disability bene…ts; for women, we will only consider the female acquaintances. We discuss the sensitivity of our results to this de…nition of the reference group variables below.
The distribution of reported DI receipt in the reference group by gender and age group is presented in Table 4. Here and in the rest of the paper we combine the categories of prevalence of DI-receipt in the reference group to three: "Nobody", "Very Few", "A Few/Many", because the frequencies for "Few" and particularly "Many" are small. Young people typically know no one on disability bene…ts. The number of reference group members on disability bene…ts is highest for 55-64 year old respondents, who also most commonly receive disability bene…ts themselves. People
older than 65 may often have a work disability (see Table 2) but hardly ever receive disability bene…ts - they receive a government provided pension and often one or more additional occupational pensions. The number of women on disability bene…ts in women’s reference groups is typically smaller than the number of men on disability bene…ts in men’s reference groups, particularly at older ages. This may be because women in older cohorts often stopped working at an early age (e.g. to raise children) and never qualify for disability bene…ts after that.
Plausibly, these reference group variables are endogenous to the respondent’s own work disability –respondents who have a work disability will often not work and will not only receive disability bene…ts, but will also more easily get acquainted with other people on disability bene…ts. Hence we will treat the number of acquaintances on disability bene…ts as a dependent variable, modelled jointly with work limita-tions. Table 5 shows cross tabs of self-reported work limitations and self reported prevalence of DI-receipt in one’s reference group. For simplicity of presentation, we combine categories for self reported work disability to three: "Not Limited", "Mildly Limited", "Moderately Limited/Severely Limited/Extremely Limited". The table clearly illustrates a positive relation between self reported work limitations and the number of people in one’s reference group drawing disability bene…ts.
Table 3. Sample Statistics for Independent Variables Mean / percentage Stroke 1.3 Cancer 3.8 Lung disease 6.0 Heart disease 7.1
High blood pressure 19.2
Diabetes 4.8
Emotional problems 11.0
Arthritis 10.4
Problems with vision 3.8
Often pain 25.4
Age in years 47.6
Low education level 39.1
Medium education level 38.7
High education level 22.1
Female 49.9
Northern provinces* 14.3
Eastern provinces* 21.6
Western provinces* 38.7
Southern provinces* 25.5
Age in reference group <25 8.7
Age in reference group 25-35 20.2
Age in reference group 36-45 27.0
Age in reference group 46-55 19.7
Age in reference group 56-65 14.7
Age in reference group 66+ 9.8
Low education level in the reference group 24.9 Medium education level in the reference group 47.9 High education level in the reference group 27.2 Notes: Data are weighted. Estimation sample N=1764.
of observations for which the dummy has value 1.
*Northern provinces are Groningen, Friesland & Drenthe *Eastern provinces are Overijssel, Flevoland & Gelderland
*Western provinces are Utrecht, Noord-Holland & Zuid-Holland *Southern provinces are Zeeland, Noord-Brabant & Limburg All variables other than "Age in years" are dummies. The table
Table 4. Distribution of DI Receipt in Reference Group by Age, % 15-24 25-34 35-44 45-54 55-64 65+ Total None 82.9 65.6 52.5 55.1 39.4 53.8 56.7 Very few 17.1 31.5 41.5 36.6 44.1 34.7 35.5 A few/many 0 2.9 5.9 8.4 16.5 11.4 7.8 No of observations 29 174 221 248 196 199 1067 15-24 25-34 35-44 45-54 55-64 65+ Total None 76.4 67.8 60.7 62.6 58.9 55.2 62.6 Very few 23.6 29.0 35.7 30.4 32.9 38.2 32.4 A few/many 0 3.2 3.6 7.1 8.2 6.5 5.0 No of observations 39 188 217 212 140 117 913
Notes: Data are weighted. Complete sample N=1980.
Women, Age Group Men, Age Group
Table 5. Self-Reported Work Disability and DI Receipt in Reference Group DI Receipt in the Reference Group, %
None Very few
A few / many Total 63.4 32.2 4.4 100.0 68.1 58.1 38.0 62.5 51.9 38.4 9.7 100.0 20.8 25.9 30.9 23.3 45.3 38.7 16.0 100.0 11.1 16.0 31.1 14.2 Total 58.1 34.6 7.3 100.0 100.0 100.0 100.0 100.0
Notes: Data are weighted. Estimation sample N=1764.
3
A Model with Reference Groups
Our econometric model explains the reported number of people on disability in the reference group R (cf. Table 4), self-reported work disability Y (cf. Table 2), and disability of vignette persons Y1; : : : ; Y15 (cf. Table 1).
Self-reports of own work disability
Individuals evaluate the extent of their work disability with a self-evaluation of whether their health problems and working conditions are su¢ ciently problematic to place them above their own subjective threshold of being somewhat limited or more than somewhat limited. The result of that evaluation depends on the extent of their true health problems as well as their subjective thresholds of what constitutes a disability, both of which vary across individuals.
More formally, self-reported work disability Y of respondent i is modeled on a 3-point scale of not at all limited, somewhat limited, and more than somewhat limited (combining the three most serious categories "moderate," "severe," and "extreme," to one) as follows:
Yi = Xi + i (1)
Yi = j if j 1i < Yi j
i; j = 1; 2; 3: (2)
For notational convenience, we de…ne 0i = 1 and 3i =1. The remaining
thresh-olds 1
i and 2i will be modeled as functions of observable and unobservable
respon-dent characteristics as describe below. The error term i is assumed to be standard
normally distributed.
Since thresholds depend on respondent characteristics, self-reported work disabil-ity alone is not enough to distinguish between variation in Yi and variation in the thresholds, i.e., genuine variation in work related health and variation in what con-stitutes a disability in respondents’perceptions. The value of vignettes is that they help to identify this distinction.
Vignette evaluations
The vignettes provide all respondents with the descriptions of the same set of work disability problems. As a consequence, variation in how respondents evaluate the given health problems informs us about variation in the subjective thresholds used by the respondents. More formally, the evaluations Yl
i of vignettes l, l = 1; : : : ; 15,
are given by
Yil = j if j 1i < Yil ji; j = 1; 2; 3: (4) Here Fl
i is a dummy variable indicating whether the person described in the
vignette is female (Fl
i = 1) or not (Fil = 0). This speci…cation follows earlier work
by Kapteyn et al. (2006), who …nd that respondents (both males and females) tend to be "harsher" on female than on male vignette persons, i.e., < 0. We assume that all l
i are independent of each other and of the other error terms, and follow
a normal distribution with mean zero and variance 2
v. Thus the li are interpreted
as idiosyncratic noise driving vignette evaluations; they re‡ect arbitrariness in each separate evaluation. If respondents have a persistent tendency to give low or high evaluations, this will not be captured by l
i but by an unobserved heterogeneity term
in the response scales, see below.
Response scale thresholds
The critical identi…cation assumption is that individuals use the same scales in eval-uating themselves as they do with the vignette persons (response consistency, cf. King et al. (2004). The thresholds used in the vignette evaluation can vary across all types of individual attributes. In this research, we expand those attributes to include the number of persons within the circle of friends who are on disability Ri. The thresholds 1
i and 2i are modeled as follows: 1 i = Vi 1+ R1Ri + i (5) 2 i = 1 i + e Vi 2+ R2Ri (6)
We have included a vector Vi of respondent characteristics (independent of all error
terms and not including the reference group variables) to allow for a rather general way in which response scales vary with individual characteristics. The distance between the two thresholds is also allowed to depend on these characteristics. The exponential forces it to be positive, as in King et al. (2004). The key parameters of interest are R
1 and R2, the estimated impact of the number of people on DI in
one’s reference group on the threshold that is used to evaluate work disability. In particular, R
1 is expected to be negative: people who know many people on disability
bene…ts will think of work disability as something common and will more often evaluate people (including themselves) as work disabled, thus using lower thresholds.4 The term i re‡ects unobserved heterogeneity in thresholds. For computational convenience, we assume that that there is no unobserved heterogeneity in the distance
4In the empirical work, we will allow the parameters R
1 and R2 to depend on education level,
between the two thresholds. i is assumed to follow a normal distribution with
variance 2, independent of Vi, Xi and XiR, and of the other error terms in the
model, i, 1i, ..., 15i and !Ri (introduced below).
Disability in the reference group
As explained above, we consider disability in the respondent’s reference group of the respondent’s own sex and combine the outcomes "few" and "many" because of the small number of observations with the latter outcome. Thus we obtain an ordered response variable with three possible outcomes, j = 1 ("none"), j = 2 ("very few") and j = 3 ("a few" or "many"). This will be modeled with an ordered probit equation: Ri = XiR R+ !iR; !Ri N (0; 2!) (7) Ri = j if j 1<R i j i; j=1;2;3:(8)
For notational convenience, we de…ne 0i = 1 and 3i = 1. Below we will further specify the thresholds 1i and
2
i. The vector XiRof respondent characteristics
driving reference group disability is assumed to be independent of all the errors in the model. Equation (7) has a "reduced form" nature in the sense that we do not explicitly model how work disability and labor force status a¤ect disability in the reference group. The exogenous determinants of labor force status and disability are included among the regressors XR
i to account for this.
Since it is likely that there are common unobserved factors a¤ecting both the number of people one knows on work disability and one’s own evaluation of work disability, we allow for correlation between i and !Ri : This correlation also allows
for the role of actual labor force status (which is not included explicitly in the model but "substituted out"): work disability drives labor force status, and labor force status drives the composition of the reference group. We will assume that i and !Ri
follow a bivariate normal distribution with correlation coe¢ cient . The vector Xi
of respondent characteristics driving work disability is assumed to be independent of all the errors in the model.
We allow for a common unobserved heterogeneity component driving the thresh-olds ji; j = 1; 2 and the thresholds in the reference group equation ki; k = 1; 2 by specifying: 1i = 10+ i and i2 = 20+ i. We normalize 10 = 0. The parameter could be positive (respondents exaggerating their work disability also exaggerate their number of acquaintances on DI) or negative (respondents who think of work
disability as something exceptional will tend to interpret a given number of acquain-tances on DI as large).
The additional parameters to be estimated are 02 and . De…ne uR
i = !Ri i
and, for notational convenience, let 00 = 1 and 30 = 1. For normalization we set Var(uRi ) = 1: We can then rewrite (8) as
Ri = j if j 10 < X R i R+ uR i j 0; j = 1; 2; 3: (9)
The assumption that i is independent of i implies that people with higher
thresholds do not tend to have larger or smaller genuine work disability (on a contin-uous scale), keeping observed characteristics Xi and Vi constant. This assumption
helps in identi…cation and avoids the need for exclusion restrictions. The assump-tion seems quite plausible, although one might argue that lower thresholds point at unobserved characteristics such as pessimistic views that can also genuinely reduce respondents’ability to work.
4
Results
We estimate the models using simulated maximum likelihood. Appendix C contains the likelihood according to the model formulated above. The integrals in the likeli-hood contributions ((15) in Appendix C) were replaced by smooth simulation-based approximations, by drawing 200 times from the joint distribution of and uR and
using Halton draws.5 Experiments with a substantially larger number of draws did not lead to appreciable di¤erences in the results, implying that the number of draws is large enough to provide an accurate approximation of the integral.
4.1
Estimation results
Table 6 presents the estimation results for self reported work disability and for DI receipt in the reference group (equations (1) and (7)). The estimates for the threshold equations (5) and (6) are given in Table 7. Estimates for the vignette equations (3) are not of primary interest and are therefore presented and discussed in Table D.1 in Appendix D.
5We have used the program mdraws written by Lorenzo Cappellari and Stephen P. Jenkins. See
Coef. s.e. Coef. s.e. age 0.290 0.146 0.326 0.170 age squared -0.021 0.014 -0.020 0.016 medium education -0.063 0.084 0.076 0.076 higher education -0.273 0.085 -0.049 0.083 female 0.026 0.070 -0.339 0.063 stroke 1.304 0.340 0.070 0.130 cancer 0.337 0.146 0.136 0.153 lung 0.617 0.133 0.185 0.172 heart problems 0.872 0.131 0.022 0.182 highblood 0.037 0.084 -0.166 0.201 diabetes 0.403 0.178 0.006 0.062 emotional problems 0.625 0.097 -0.029 0.075 arthritis 0.423 0.118 -0.016 0.081 vision 0.098 0.178 0.019 0.071 often pain 1.231 0.081 -0.158 0.064 intercept -2.005 0.376 -0.064 0.257
reference group age 25-35 -0.177 0.160
reference group age 36-45 0.295 0.134
reference group age 46-55 0.003 0.119
reference group age 56-65 0.065 0.075
reference group age >65 0.155 0.153
Medium education in R.G. 0.268 0.099 High education in R.G. 0.197 0.109 northern provinces 0.041 0.166 eastern provinces 0.252 0.078 western provinces -1.250 0.388 0.055 0.039 1.336 0.051
Table 6. Estimation Results for Own Work Disability and DI Receipt in Reference Group Own Work Diability DI Receipt Ref. group
ρ
Work disability self-reports
The equation for self-reported disability in Table 6 shows that self-reported disability goes up with age until age 69; it is lower for higher educated individuals and higher for individuals with serious health conditions, including strokes, heart problems, cancer, diabetes, emotional problems, and lung problems.
4.1.1 DI receipt in the reference group
The reference group DI receipt equation shows that the reported DI prevalence in the reference group increases with age (the top of the parabola is estimated at 82 years); it shows virtually no relation with education, and implies that DI receipt in the reference group increases with most health conditions, in line with the argument that people with a health problem will more often be acquainted with other people in poor health. Also in line with the raw data (Table 4) is that females are less likely to report to have DI-bene…t recipients in their reference group. Respondents in the western provinces are less likely to know people on disability bene…ts than respondents in the rest of the country.
The variables that a¤ect the number of people on DI in the reference group are of interest in part because, as we shall see below, the number of people in the reference group signi…cantly a¤ects the thresholds used in evaluating work disability. For example, women know fewer people on DI and because of that will less easily say that a given health problem constitutes a work disability. Similarly, having pain increases the number of people on DI in one’s reference group this makes people with pain ’softer’in evaluating disability.
Thresholds
The results for the threshold equations are presented in Table 7. The top panel presents estimates for the coe¢ cients on individual characteristics in equations (1) and (7), while the bottom part shows estimates of the coe¢ cients of peer group DI receipt Ri interacted with education, age, and gender in both threshold equations. The mean value of Ri is slightly positive. Thus the estimates for the …rst threshold
Table 7. Estimation Results Threshold Equations
Coef. s.e. Coef. s.e.
age 0.234 0.098 -0.076 0.028 age squared -0.019 0.009 0.007 0.003 medium education 0.071 0.039 -0.045 0.016 higher education 0.080 0.038 -0.032 0.015 female -0.110 0.051 0.021 0.014 stroke -0.100 0.154 -0.051 0.053 cancer -0.061 0.086 0.023 0.031 lung -0.056 0.076 0.031 0.028 heart problems 0.056 0.062 -0.062 0.028 highblood 0.003 0.039 0.020 0.016 diabetes 0.000 0.079 0.025 0.032 emotional problems 0.001 0.059 -0.023 0.021 arthritis 0.039 0.059 -0.016 0.022 vision -0.057 0.082 0.031 0.036 often pain 0.100 0.048 -0.014 0.015 intercept -0.837 0.276 0.178 0.085 -0.816 0.086 0.523 0.059
Coef. s.e. Coef. s.e.
The age function has a top at about 62 years, implying that until age 62, older people are "tougher", i.e. less likely to call a condition work disabling.6 The only signi…cant health condition is pain - respondents who often su¤er from pain less easily evaluate a given health problem as a (mild or worse) work limitation.
For the distance between the …rst and second threshold ( 2), results are di¤erent;
the age function has a minimum at 54 years of age (if Ri = 0), while higher edu-cation leads to a smaller distance between thresholds. Heart problems do the same; theyare the only type of health problems with a signi…cant e¤ect. The estimates are di¢ cult to interpret individually, due to the complexity of the model, where the same variables appear in several equations. The parameters of primary interest are
R
1 and R2. Both have been parameterized as a function of education level, age, and
gender (see the bottom panel of Table 7). Consider …rst the estimated main e¤ect and the interactions with education. For males under 35 with lower education R
1 is
estimated at .45; for the same individuals with medium education the estimate is -.42 (not signi…cantly di¤erent from the -0.45 estimate), while for the higher educated the estimate is -.33. The di¤erence between the highest and lowest education groups is statistically signi…cant. Females are signi…cantly less in‡uenced by DI receipt in the reference group than males, although quantitatively the di¤erence is not big. Interactions of DI receipt in the reference group with age dummies are small and insigni…cant.
As long as the estimated value of R
1 is negative (as it is in all cases), the fraction
of people who are on DI bene…ts in the reference group will unambiguously shift the reporting thresholds for own disability. In this sense, R1 is the more critical
parameter of the two. The estimates for R
2 show that the distance between thresholds increases with the
number of acquaintances on disability bene…ts, particularly for the lowest education level and age less than 35 or above 65. In simulations, we …nd that the e¤ects of increasing numbers of people on DI in the reference group results both in increases in the fraction of those reporting they are somewhat limited and those reporting they are moderately limited or worse.
As mentioned earlier, we de…ned reference groups separately for men and women in the sense that for women we took the number of women on DI amongst female acquaintances and for men the number of male DI recipients among male acquain-tances. One question is how sensitive our results in Table 7 are to this particular formulation of reference groups. To test this, we re-estimated the model using a
mon de…nition of reference groups for both sexes.7 The estimated e¤ects of number
of people on DI in the reference group are even larger using the common reference by gender than with the benchmark de…nition used for Table 7. A log likelihood test however indicates that the model we estimate with separate reference groups by gender is signi…cantly better that the alternative model.
4.1.2 Covariance structure of the errors
Table 6 shows that the parameter , the correlation between the error terms in the equations for own work disability (7) and DI receipt in the reference group (1), is small and insigni…cant.
Unobserved heterogeneity in thresholds is signi…cant - the estimated standard deviation of is 0.52 and seems to be very accurately determined ( in Table 7). To judge its size, it can be compared to the amount of idiosyncratic noise in self-reports and vignette evaluations. The former has standard deviation 1 (by normalization), the latter has standard deviation 0.51 (see Table D.1). Thus unobserved heterogene-ity in thresholds explains about 23% of the unsystematic variation in self-reports and about 53% of the unsystematic variation in vignette evaluations.
The parameter is estimated at -.82. Since uR = !R and Var(uR
i )= 1, we
have Var(!R
i )=.82. The implied correlation between i and uRi is equal to -.43. The
sign of implies that if a respondent uses high thresholds for answering about his or her own work limitations, he or she will tend to use low thresholds when asked for DI prevalence in the reference group. Thus someone who is unlikely to refer to a health problem as work limiting, will sooner consider a given number of people on DI in the reference group as "many".
Model performance
Table 8 provides a simple way of checking the …t of the model. It is similar to Ta-ble 5, but reports frequencies simulated using the model. Comparing TaTa-ble 8 with Table 5 suggests that the …t of the model is fairly good; judging from the marginal distributions, the model does a good job in replicating reported reference group DI-receipt; it does a slightly worse job in reproducing the distribution of self-reported disability. The biggest deviation between the data and the model predictions occurs
7All respondents were asked both the numbers of men and the numbers of women on DI in the
reference group. To form a common de…nition, we used the maximum of the two. Thus if for an individual respondent there were a lot of individuals of one gender who were more than somewhat limited, that is the value that applies.
in the middle category (mildly limited). According to the data, 23.3% of the respon-dents classify themselves as mildly limited, whereas the model predicts 19.6% in that category.
Disability in the reference group, % None Very few
A few / many Total Not limited 62.1 32.3 5.7 100.0 71.8 60.4 48.7 66.0 Mildly limited 51.4 38.4 10.3 100.0 17.6 21.3 26.2 19.6 Moderately, severely
and extremely limited 42.1 44.5 13.4 100.0
10.6 18.2 25.1 14.4
Total 57.1 35.2 7.7 100.0
100.0 100.0 100.0 100.0
Notes: Data are weighted. Estimation sample N=1764.
Table 8. Model Predictions of Self-Reported Work Disability and DI Receipt in Reference Group
Self-re port ed work disa bility , %
4.2
Simulation of reference group e¤ects
One way to gauge the strength of the reference group e¤ects is to arti…cially vary the number of people on DI in an individual’s reference group and then to evaluate how this a¤ects the prevalence of self reported work limitations. We do this by varying the intercept in the equation for the number of people in the reference group on DI (7) and then simulate the reports of DI-bene…t receipt in the reference group and the prevalence of self reported work disability induced by that new level of reference group DI-receipt.
(sub)sample percentage (except the lower line in the …rst …gure). The curves in the …gures illustrate the sensitivity of reporting a work disability for DI receipt in the reference group. In line with the estimation results in Table 7, the slope is largest for the low educated and smallest for those with high education level. But in all cases there is a notable e¤ect - if the respondent knows more people on DI bene…ts, the chances of reporting a disability increase substantially.
To illustrate the size of the e¤ect, in the picture for the full sample, a second horizontal line and vertical line have been drawn, both below the sample averages. The horizontal line is based on the …nding of Kapteyn et al. (2006). that if US scales are assigned to Dutch respondents, self-reported work limitations in the Netherlands would fall by 21%.8 This second horizontal can thus be interpreted as self-reported work-limitations in the Netherlands if US scales are applied to Dutch responses. The second vertical line shows that if the percentage of individuals saying they know at least a few DI-bene…t recipients in their reference group were to move from its simulated sample mean of 42.9% to about 18.2% (the left most vertical line), this would move the scales used by the respondents enough to reach the US scales.
The level of the graphs for the three education groups varies quite a bit, with the graph for the lowest education group being the highest. That is, at the same level of perceived reference goup DI bene…t receipt lower educated respondents are more likely to report at least a mild work limitation than respondents with middle or higher education.
5
Concluding Remarks
Most people do not live in social isolation. Instead, they interact repeatedly with family, friends, and neighbors. As a consequence of those pervasive interactions, they allow themselves to be transformed in many ways, a transformation of which they may often be unaware. One type of transformation involves the formation of social norms about what normal or acceptable behavior might be. These social norms then …x the thresholds that they may be using in responding to questions about their own behaviors and current situations. If they had di¤erent neighbors and friends, their self-descriptions about their lives may well be quite di¤erent. While this may be true within a country where there exists a shared history and culture, it is especially likely to be the case when cross-national comparisons are made.
8According to their benchmark model; the percentage varies slightly depending on which model
speci…cation is chosen.
0 20 40 60 80 Pe rc e n t m ild o r more 0 20 40 60 80 100
Percent knowing at least a few DI-recipients Complete sample 0 20 40 60 80 Pe rc e n t m ild o r more 0 20 40 60 80 100
Percent knowing at least a few DI-recipients Low education 0 20 40 60 80 Pe rc e n t m ild o r more 0 20 40 60 80 100
Percent knowing at least a few DI-recipients Median education 0 20 40 60 80 Pe rc e n t m ild o r more 0 20 40 60 80 100
Percent knowing at least a few DI-recipients High education
Fig. 1: Self-reported work disability and reference group DI
In this paper, we test the importance of these types of social interactions using a speci…c application- the probability that people self-label themselves as work dis-abled. We estimated a model of self reported disability with an emphasis on how the reporting of disability is a¤ected by the prevalence of DI receipt in one’s reference group. We …nd an e¤ect in the hypothesized direction- larger reported numbers of people in one’s reference group on DI increases the likelihood of seeing oneself as having a work disabilty.
These …ndings are suggestive of how policy programs a¤ect social norms. If a policy makes receipt of DI bene…ts more attractive or easier (e.g., by loosening eligibility requirements) thus increasing the number of DI recipients, this changes social norms. Individuals are now more likely to label a given health condition as work limiting and the prevalence of self-reported work will rise.
by itself will generate correlation between self-reported disability and reported DI bene…t receipt in one’s reference group. But we also allow for correlation between the errors in the reference group equation and the equation predicting the probability that one is work disabled. Even within this reasonably general model, we …nd a direct e¤ect of the number of people in one’s reference group on disability programs on the probability one considers oneself work disabled.
The e¤ects that we estimate are su¢ ciently strong that they are able to explain a good deal of the higher rates of work disability in the Netherlands compared to the United States. The Dutch population appears to have much more lenient thresholds about what constitutes a work disability (Kapteyn et al. 2006). The results in this paper suggest that this tendency stems from the fact that the Dutch are much more likely to know people on work disability programs, a direct consequence of the far more generous programs in The Netherlands as well as its more lenient rules for program eligibility. Figure 1 suggests that for our reference group hypothesis to explain the complete di¤erence, Americans should be about half as likely to report that they know at least a few people in their reference group receiving DI bene…ts. It is suggestive that the actual prevalence of DI bene…t receipt in the U.S. is a little less than half of what it is in The Netherlands.
References
Aarts, L., Burkhauser, R. & de Jong, P. (1996), Curing the Dutch Disease: An International Perspective on Disability Policy Reform, Advanced Textbooks in Economics, Aldershot Avebury.
Alessie, R. & Kapteyn, A. (1991), ‘Habit formation, interdependent preferences and demographic e¤ects in the almost ideal demand system’, Economic Journal 101, 404–419.
Audretsch, D. & Feldman, M. (1996), ‘R and d spillovers and the geography of innovation and production’, American Economic Review 86, 630–640.
Bala, V. & Goyal, S. (1998), ‘Learning from neighbours’, Review of Economic Studies 65, 595–622.
Banerjee, A. V. (1992), ‘A simple model of herd behavior’, Quarterly Journal of Economics 107, 797–817.
Banks, J., Kapteyn, A., Smith, J. P. & van Soest, A. (2005), ‘Work disability is a pain in the *****, especially in england, the netherlands, and the united states’, RAND Labor and Population Working Paper WR-280.
Becker, G. S. (1974), ‘A theory of social interactions’, Journal of Political Economy 83(6), 1063–1093.
Bound, J. (1991), ‘Self-reported versus objective measures of health in retirement models’, Journal of Human Resources 26(1), 106–138.
Bound, J. & Burkhauser, R. (1999), Economic analysis of transfer programs targeted on people with disabilities, in O. Ashenfelter & D. Card, eds, ‘Handbook of Labor Economics’, Vol. 3C, Elsevier, Amsterdam, pp. 3417–3528.
Cappellari, L. & Jenkins, S. (2006), Calculation of multivariate normal probabilities by simulation, with applications to maximum simulated likelihood estimation. Discussion Paper No. 2112, IZA, Bonn.
Case, A. & Katz, L. (1991), The company you keep: The e¤ects of family and neighborhood on disadvantaged youth. NBER Working Paper 3705, National Bureau of Economic Research, Cambridge, MA.
Duesenberry, J. (1949), Income, Saving and the Theory of Consumer Behavior, Har-vard University Press.
Du‡o, E. & Saez, E. (2003), ‘The role of information and social interactions in retirement plan decisions: evidence from a randomized experiment’, Quarterly Journal of Economics 118(3), 815–842.
Ginther, D., Haveman, R. & Wolfe, B. (2000), ‘Neighborhood attributes as deter-minants of children’s outcomes: How robust are the relationships’, Journal of Human Resources 35(4), 603–642.
Glaeser, E. L., Sacerdote, B. & Scheinkman, J. A. (1996), ‘Crime and social interac-tions’, Quarterly Journal of Economics 111(2), 507–548.
Glaeser, E. L., Sacerdote, B. & Scheinkman, J. A. (2000), ‘The social multiplier’, Journal of the European Economic Association 1(2-3), 345–353.
Haveman, R. & Wolfe, B. (2000), ‘The economics of disability and disability pol-icy’, Handbook of Health Economics Vol. 1B J. Newhouse and A. Culyer (eds) pp. 995–1051.
Kapteyn, A., Smith, J. P. & van Soest, A. H. (2006), ‘Vignettes and self-reports of work disability in the us and the netherlands’, American Economic Review forthcoming.
Kapteyn, A., van de Geer, S., van de Stadt, H. & Wansbeek, T. (1997), ‘Interdepen-dent preferences: An econometric analysis’, Journal of Applied Econometrics 12(6), 665–686.
King, G., Murray, C., Salomon, J. & Tandon, A. (2004), ‘Enhancing the validity and cross-cultural comparability of measurement in survey research’, American Political Science Review 98(1), 567–583.
Manski, C. F. (1993), ‘Identi…cation of endogenous social e¤ects’, The Review of Economic Studies 60(3), 531–542.
Topa, G. (2001), ‘Social interactions, local spillovers and unemployment’, The Review of Economic Studies 68(2), 261–295.
Veblen, T. (1899), The Theory of the Leisure Class, Dover.
Woittiez, I. & Kapteyn, A. (1998), ‘Social interactions and habit formation in a model of female labour supply’, Journal of Public Economics 70, 185–205.
A
Vignette Questions
Vignettes for A¤ect
1. [Henriette] generally enjoys her work. She gets depressed every 3 weeks for a day or two and loses interest in what she usually enjoys but is able to carry on with her day-to-day activities on the job.
2. [Jim] enjoys work very much. He feels that he is doing a very good job and is optimistic about the future.
3. [Tamara] has mood swings on the job. When she gets depressed, everything she does at work is an e¤ort for her and she no longer enjoys her usual activities at work. These mood swings are not predictable and occur two or three times during a month.
4. [Eva] feels worried all the time. She gets depressed once a week at work for a couple of days in a row, thinking about what could go wrong and that her boss will disapprove of her condition. But she is able to come out of this mood if she concentrates on something else.
5. [Roberta] feels depressed most of the time. She weeps frequently at work and feels hopeless about the future. She feels that she has become a burden to her co-workers and that she would be better dead.
Vignettes for Pain
1. [Katie] occasionally feels back pain at work, but this has not happened for the last several months now. If she feels back pain, it typically lasts only for a few days.
2. [Catherine] su¤ers from back pain that causes sti¤ness in her back especially at work but is relieved with low doses of medication. She does not have any pains other than this generalized discomfort.
3. [Yvonne] has almost constant pain in her back and this sometimes prevents her from doing her work.
5. [Mark] has pain in his back and legs, and the pain is present almost all the time. It gets worse while he is working. Although medication helps, he feels uncomfortable when moving around , holding and lifting things at work.
Vignettes for CVD
1. [Trish] is very active and …t. She takes aerobic classes 3 times a week. Her job is not physically demanding, but sometimes a little stressful.
2. [Norbert] has had heart problems in the past and he has been told to watch his cholesterol level. Sometimes if he feels stressed at work he feels pain in his chest and occasionally in his arms.
3. [Paul]’s family has a history of heart problems. His father died of a heart attack when Paul was still very young. The doctors have told Paul that he is at severe risk of having a serious heart attack himself and that he should avoid strenuous physical activity or stress. His work is sedentary, but he frequently has to meet strict deadlines, which adds considerable pressure to his job. He sometimes feels severe pain in chest and arms, and su¤ers from dizziness, fainting, sweating, nausea or shortness of breath
4. [Tom] has been diagnosed with high blood pressure. His blood pressure goes up quickly if he feels under stress. Tom does not exercise much and is overweight. His job is not physically demanding, but sometimes it can be hectic. He does not get along with his boss very well.
5. [Dan] has undergone triple bypass heart surgery. He is a heavy smoker and still experiences severe chest pain sometimes. His job does not involve heavy physical demands, but sometimes at work he experiences dizzy spells and chest pain.
B
Reference Group Questions
The questions are preceded by the following introduction:
The following questions concern your circle of acquaintances, that is, the people with whom you associate frequently, such as friends, neighbors, acquaintances, or maybe people at work.
If you think of your circle of acquaintances, into which age category do MOST of these people go? Please select the answer that is closest to reality.
age (in years) is mostly: 1 under 16 2 16 - 20 3 21 - 25 4 26 - 30 5 31 - 35 6 36 - 40 7 41 - 45 8 46 - 50 9 51 - 55 10 56 - 60 11 61 - 65 12 66 - 70 13 71 or over
Which level of education do most of your acquaintances have? 1 primary education
2 junior vocational training 3 lower secondary education
4 secondary education/pre-university education 5 senior vocational training
6 vocational colleges/…rst year university education 7 university education
1 Nobody 2 Very few 3 A few 4 Many
C
Likelihood Contributions
Compared to the models in King et al. (2004) and Kapteyn, Smith, and Van Soest (2006), there are two complications: the thresholds now depend on an unobserved variable R and upon an unobserved heterogeneity term . Replacing R using (7) and exploiting (5) and (6) gives:
1 = V 1+ R 1X R R + + R1(uR+ ); (10) 2 = 1+ eV 2+ R 2XR R+ R2(uR+ ) (11) (1) and (2) imply Y = j if j 1 X < < j X (12)
Similarly, for the vignette evaluations we get:
Yl = j if j 1 l Fl < l< j l Fl (13)
The probability of observing a certain reference group category follows from (9): R = j if 0;j 1 XR R < uR < 0j XR R (14) Let the reported reference group variable be r, the observed work disability self-report y, and the observed vignette evaluations y1; : : : ; yL. Then the likelihood
con-tribution of a given respondent can be written as a two-dimensional integral over the values of uR that result in R = r and all possible values of :
Z 1 1 Z 0j XR R 0;j 1 XR R P (Y = yjuR; ) L Y l=1 P (Yl = yljuR; )f (uRj )duR 1 ( )d (15) where is the standard normal density and f is the conditional density of uR given
, which is univariate normal. Of course, the crucial point here is that, conditional on uR and , all vignette evaluations and the self-report are mutually independent,
allowing for the factorization in (15). The conditional probabilities in (15) follow from (12) and (13), together with the normality assumptions on the error terms, implying that the l are independent of , and uR
but that j(uR; ) N ( uR; 1 2):
P (Y = yjuR; ) = ([ y X uR]=p[1 2])
P (Yl= yljuR; ) = ([ yl l Fl]= v)
([ yl 1 l Fl]= v)
where the ::: are given by (10) and (11) (and depend on and uR).
D
Estimates of Vignette Dummies
The dummy coe¢ cients in Table D.1 re‡ect the average severity of the work lim-itations described in the vignettes. One can relate the dummy coe¢ cients to the relative frequencies in Table 1. The estimate of , the coe¢ cient of the dummy for a female vignette name is small and insigni…cant. The estimated idiosyncratic vari-ation in vignette evaluvari-ations v (independent across vignettes) is smaller than the
unsystematic variation in self-assessments ( = 1, by means of normalization).
Vignette Coef. s.e.
1 0.000 2 -1.273 0.065 3 0.705 0.040 4 0.532 0.031 5 1.228 0.064 6 0.297 0.024 7 0.746 0.038 8 1.566 0.077 9 1.527 0.076 10 1.340 0.067 11 -0.945 0.048 12 0.674 0.036 13 1.132 0.058 14 0.479 0.029 15 0.782 0.043 0.507 0.023 -0.007 0.019
Table D.1. Estimates of Vignette Equations
